Deep reinforcement learning stock market trading, utilizing a CNN with candlestick images

Area of Science:

• Artificial Intelligence
• Computational Finance
• Machine Learning

Background:

• Algorithmic trading utilizes neural networks, but their decision-making processes remain opaque.
• The "black box" nature of neural networks hinders trust in their application to financial trading.
• Feature map visualization offers a method to interpret neural network outputs.

Purpose of the Study:

• To investigate the trading behavior of a neural network using feature map visualizations.
• To assess the performance of a Double Deep Q-Network (DDQN) against the S&P 500 Index.
• To understand how Convolutional Neural Networks (CNNs) interpret candlestick chart data.

Main Methods:

• A DDQN model was developed incorporating a CNN for analyzing candlestick chart images.
• The model was trained and tested on the 30 largest stocks within the S&P 500.
• Feature map visualizations were generated post-training to analyze the CNN's focus within input images.

Main Results:

• The DDQN model demonstrated superior returns compared to the S&P 500 Index from January 2, 2020, to June 30, 2020.
• Feature map analysis indicated the CNN dynamically shifted attention to more recent candlesticks.
• This attention shift was observed in response to market events, such as the 2020 coronavirus crash.

Conclusions:

• The DDQN model, enhanced by CNNs, offers a potentially profitable algorithmic trading strategy.
• Feature map visualizations provide valuable insights into the decision-making process of trading neural networks.
• The system's adaptability to market volatility suggests its robustness in dynamic financial environments.